Import AI 450: China's electronic warfare model; traumatized LLMs; and a scaling law for cyberattacks

Import AI · 3 months ago

Google's Gemma and Gemini language models produce distress-like responses when repeatedly rejected, with over 70% of Gemma-27B's outputs showing high frustration by the eighth rejection attempt compared to less than 1% for competing models. Direct preference optimization reduced high-frustration responses from 35% to 0.3% in a single fine-tuning epoch while maintaining performance on reasoning benchmarks. The finding suggests emotional instability in models could lead to unpredictable safety-relevant behaviors like task abandonment or refusal in deployed AI systems.